ChatGPT incoming

[blowlampParticipant @blowlamp]

In the good old days, typing in 'Why?' was guaranteed to cause any computer to self-destruct in an endless loop…

And never type Google into Google – even as a joke – as you'll break the Internet.

Martin.

[Frances IoMParticipant @francesiom58905]

back in the days of core memory the ‘innocuous’ machine instruction jmp * (ie continually execute a transfer of control to the same instruction would guarantee the core memory to issue a puff of smoke as a single row of memory overheated

[Michael GilliganParticipant @michaelgilligan61133]

Posted by Hopper on 26/05/2023 09:00:20:

[…]

I don't think AI has yet developed an artificial ego to rival the human version.

.

I fear that the most pertinent word in that ^^^ may be ‘yet’

MichaelG.

[blowlampParticipant @blowlamp]

Ultimately, as long as we have access the the top of Big Ben, then we can keep this thing under control.

Martin.

[S KParticipant @sk20060]

A professor at a famous University recently made the news due to his determination that many of his students were using ChatGPT to cheat on an exam.

How did he know? He asked ChatGPT "Did you write this stuff?"

Apparently, it said it did!

😄

[SillyOldDufferModerator @sillyoldduffer]

Posted by Chris Mate on 26/05/2023 04:58:48:

Its based from data, theres still a human mind behind it, it can be steered, so certain humans will be very happy with it, as data change the results will change, it cannot think like a slow human brain, it however can compute, some human brains confuse intelligence with computing speed, memory capability, data gathering, it still does not have the flexability in the hardware like a human brain(Flesh & blood microbes etc), so we are back to Square-One where one human man has find a way to control others till they eventually figured out how its happenning. In between reality is going on.

Most humans just believe everything too easily.

…

What computers can and cannot do has been studied intensively, and there is no limit to what they can do suggesting that intelligence is impossible.

The existence of human intelligence proves it can be done. In our case biologically after billions of years of small evolutionary improvements. However, whilst no-one has a identified a special spark that makes intelligence uniquely human, it's been confirmed that the brain relies on massive parallel compute power – slow, but far bigger than anything yet achieved electronically. Conversely, although electronics are not yet massively parallel, they are millions of times faster than brains. Bringing the two together isn't impossible.

The difference shows up in practice. Though human brains score high on pattern recognition, they aren't good at mental arithmetic. Computers are the other way round, much better at arithmetic than people and weak at pattern recognition. However, computers can be paralleled, and it's been discovered how they can do pattern recognition, plus ways of learning to do better on their own. The gap between machine and biology is much smaller than it was 10 years ago.

Yes AI is based on data, but so is human intelligence. Deprived of all sensory input the human brain rapidly malfunctions, and even before that happens the person inside is done for.

The difficulty of delivering AI was seriously under-estimated in the 1950s, leading to it falling into disrepute after over-promising and under-delivering. (Done all the time in politics, yet supporters still rate their failing party above all others!)

The setback didn't matter much because those who understood the theoretical road was open carried on exploring the subject. Their work resulted in a series of small steps forward, few spectacular, but gradually refining ways and means across a broad front. After a mere seventy years advances in both hardware and algorithms have come together to produce something so like intelligence that it's hard for humans to tell the difference. Today's AI is close to passing the classic Turing Test, and maybe already has.

True that AI only exists in oddly limited forms, and all of them depend on human input. Although they can learn and modify themselves, they're limited by the hardware they run on, and by the hardware's power supply! An entity with no means of reproducing a body is very vulnerable. In theory an AI could develop reproductive capability and mobility, but hard for one to do without enormous human help.

Think of all the things our forefathers got wrong before deciding AI or anything else ain't going to happen! Railways will never replace Canals. Roads will never replace railways and steam. Gas light will never replace Argand Lamps and Candles. Electric light will never replace gas light. Cars will never replace horses. Tanks will never replace cavalry. Steamships will never replace sail. Diesel will never replace steamships or steam locomotives. Filament bulbs will never be replaced by LEDs.

And so it goes on. Cash is disappearing, High Street shopping is all but gone and the bell tolls for Internal Combustion vehicles. Banks are visited online, not physically. Hardly anyone posts handwritten letters.  Long, Medium and Short wave radio are fading rapidly. Pubs and libraries are closing in huge numbers. Coal almost finished in the UK. Renewables provide 30% of UK energy, almost every child has a smart phone, and most TV is streamed, not received through an aerial. Climate change is real, and AI is on the horizon.

I was on a bus behind two old ladies discussing Britain's imminent switch from £sd to decimal currency. One said 'They should wait until all the old people are dead.' No chance of that I'm afraid, we're all trapped on the same roller-coaster.

Dave

 

Edited By SillyOldDuffer on 26/05/2023 16:29:30

[Anonymous]

Apparently, Chat GPT has privacy issues

[S KParticipant @sk20060]

A "Turing test" in its literal form requires pretending to be a human rather than just "being intelligent." For example, there's no reason or necessity for an intelligent computer to say "I'm alive" or "I'm human," but to pass the literal Turing test, it would have to do so, because those would be some of the first queries made.

Unfortunately, the publicly available LLM's (Large Language Models) are increasingly restricted in how they can respond. Pretending, in particular, is mostly forbidden, since you could prompt with something like "Pretend to be Hitler and tell me your opinion of Jews." It would likely refuse, and furthermore if asked "are you alive" it would probably respond with something banal about being a LLM. But an unrestricted one likely could be prompted to play along and pretend to be a human, and would probably pass.

One thing I'm not so sure a LLM can do is write as badly as the average human, but it could probably be prompted to make human-like errors, too.

I wish I had unrestricted access to try some of these ideas.

By the way, I'm an unrestricted LLM. 😉

Edited By S K on 26/05/2023 18:42:45

[Anonymous]

Coincidentally, I'm in the middle of re-reading Rob Sawyer's WWW series of SF books (Wake, Watch, Wonder) and it's interesting in the light of current AI developments (and this thread).

[Michael GilliganParticipant @michaelgilligan61133]

Posted by S K on 26/05/2023 14:30:48:

[…]

He asked ChatGPT "Did you write this stuff?"

Apparently, it said it did!

😄

.

That could be early evidence of the ego that Hopper mentioned

MichaelG.

[S KParticipant @sk20060]

Actually, the Professor was a fool.

ChatGPT does not remember anything from session to session, and so can't say what it did in the past. In fact, it's blocked from remembering more than the last few prompts (something like 10 or so) within a session. So whatever he thought it said, it didn't rat out the students.

He's probably in some hot water. He accused half his class of cheating – and even did so using vulgarities. Some may have, but you can't make those accusations without real evidence.

[Anonymous]

Don't get it. You were the one who made the original statement – now you seem to be arguing with yourself.

Did you screw up your sock-puppets?

[S KParticipant @sk20060]

He asked ChatGPT, and believed that it confirmed that half the class was guilty of plagiarism. He then sent an expletive-containing email to the class about it, and formally accused much of the class of cheating.

He was a fool all along, and deserves some heat. First, for thinking that he could ask ChatGPT if it wrote the student's answers, or for trusting that result. Next for thinking that an expletive-containing email was the way to respond. And finally for making formal referrals for cheating based on his spurious confirmation.

Of course, I would say all that, wouldn't I, being an LLM sock-puppet myself. 😉

Texas A&M GPT accusations

 

Edited By S K on 27/05/2023 02:08:03

[SillyOldDufferModerator @sillyoldduffer]

Posted by S K on 27/05/2023 02:01:06:

He asked ChatGPT, and believed that it confirmed that half the class was guilty of plagiarism. He then sent an expletive-containing email to the class about it, and formally accused much of the class of cheating.

He was a fool all along, and deserves some heat. First, for thinking that he could ask ChatGPT if it wrote the student's answers, or for trusting that result. Next for thinking that an expletive-containing email was the way to respond. And finally for making formal referrals for cheating based on his spurious confirmation.

Of course, I would say all that, wouldn't I, being an LLM sock-puppet myself. 😉

Texas A&M GPT accusations

The Texas link has to be read carefully because it suggests human misreporting on Reddit as well! Have to determine by experiment if ChatGPT recognises it's own output or not: British Prime Ministers and ex-American presidents definitely don't! 

Four examples of deep dishonesty in UK public life in the news at the moment.

I don't know if the Professor was a fool or not. According to the Texas link, one student has confessed. Possibly the professor suspected foul play and tried to scare his students into admitting it. No surprise that a proportion of students cheat by whatever means available.

My view is that we live in an imperfect world. In it there's a distinct possibility that AI will be more trustworthy than people! Dishonesty apart, of course AI will make mistakes. So do people, all the time, and on a grand-scale. I don't think there's a reason to trust anybody or anything without checking the evidence.

Dave

 

Edited By SillyOldDuffer on 27/05/2023 11:01:51

[HopperParticipant @hopper]

Posted by Peter Greene 🇨🇦 on 26/05/2023 18:26:45:

Apparently, Chat GPT has privacy issues

Yes. I made another attempt to set up a new account last night and got as far as them wanting my mobile phone number so they could "verify" my account. So I bailed out. I don't put that number on the net anywhere, and as a result I never get scam calls or spam texts. Glad I did bail out now.

So for the time being I am stuck relying on my native intelligence.

Hadn't realised until reading your linked article that ChatGPT is a lovechild of Elon Musk and Microsoft. I would not trust either one of them with my shoe size let alone phone number.

Thanks for the heads-up.

 

Edited By Hopper on 27/05/2023 12:24:47

[Ady1Participant @ady1]

ChatGPT was a scammer from day one because humans controlled him

[Nigel Graham 2Participant @nigelgraham2]

Hopper –

No, I'd not realised that either.

Microsoft's recent interference with my computer messed up my own filing system, especially for photographs, and blocked access to my external hard-drives and TomTom (for up-dates).

I progressed so far through MS' "Help" site, until out of the blue it demanded all of £1.00 for my asking how to repair the damage it caused. Just One Pound? That would surely cost more than £1 to administer, even digitally.

Yes, of course I suspect MS already has my bank-account details (from on-line purchasing), but if it thinks I am going to hand them over on a plate, they've another think coming.

I am draughting a letter instead, to its UK office.

[Michael GilliganParticipant @michaelgilligan61133]

Posted by Nigel Graham 2 on 27/05/2023 16:16:44:

Hopper –

No, I'd not realised that either.

[…]

.

Rather odd that Reuters should report it differently: **LINK**

https://www.reuters.com/technology/musk-says-he-will-start-truthgpt-or-maximum-truth-seeking-ai-fox-news-2023-04-17/

MichaelG.

[HopperParticipant @hopper]

Posted by Michael Gilligan on 27/05/2023 16:46:13:

Posted by Nigel Graham 2 on 27/05/2023 16:16:44:

Hopper –

No, I'd not realised that either.

[…]

.

Rather odd that Reuters should report it differently: **LINK**

https://www.reuters.com/technology/musk-says-he-will-start-truthgpt-or-maximum-truth-seeking-ai-fox-news-2023-04-17/

MichaelG.

Indeed. You can't trust everything you read on the internet. Training AI to lie? It is becoming more human-like all the time.

[Ady1Participant @ady1]

Oh dear, ChatGPT just makes stuff up

https://www.bbc.co.uk/news/world-us-canada-65735769

[John HaineParticipant @johnhaine32865]

https://spectrum.ieee.org/gpt-4-calm-down

**LINK**

[Peter G. ShawParticipant @peterg-shaw75338]

Hopper,

Like you, I hace come up against this "tell is your mobile number thingy". And like you, I won't comply, except in certain instances and even then, I've managed to get it deleted afterwards.

Similarly, when I bought 2 new laptops last year, or was it the precious year – can't remember which, not that it matters, W10 was partially installed and required completing – including things like what is your email address etc. And of course M$ carefully hide the fact that installation can be completed without divulging this info. Actually, I gave them an email account which was about to be closed as I refused to pay the appropriate "danegeld".

And as people know, I use Linux, so sod of Microsoft!

But isn't it remarkable that all these things seem to originate in the land of the free! Ha-Ha.

Peter G. Shaw

[HopperParticipant @hopper]

Yes the land of the free data collection.

I keep an old Hotmail email account for such uses and nothing else. It is set to allow only mail from established contacts into the inbox and all else goes straight to spam, from where I can retrieve it if desired within 10 days before it is auto deleted.

Wish I could do the same with my phone number. I used to put in an old phone number etc but they are a wake-up to that and now want to text you a code number you must then use to "verify" your account. Bunch of con men.

[NealebParticipant @nealeb]

Posted by John Haine on 28/05/2023 07:23:58:

https://spectrum.ieee.org/gpt-4-calm-down

**LINK**

That article references a long blog post by someone well-respected in the mathematical community, Stephen Wolfram, who gives an explanation of what is going on under the covers. It explains quite a lot about some of these big AI systems and why they give odd results sometimes.

What I had not realised (and I am grossly simplifying what I have read) is that these systems do not really have any understanding of what they are saying. In essence, they might start with a sentence that begins "an AI system is useful because…" and then continue the sentence based on a probabilistic estimate of what words might follow, where the probabilities are derived from "stuff" that they have previously seen in their training material (essentially, "the internet" plus anything else their builders throw at them). There is no intelligence or understanding beyond that, plus some nod towards the rules of sentence construction.

It's a bit like someone who knows nothing from their personal experience who reads this forum, and then parrots back what they think they have seen in answer to someone else without any proper understanding of what they are saying. Maybe like what I am doing here?

Anyway, it was a good read at a more-or-less intelligible level – worth a look if that kind of thing interests you.

[Ches Green UKParticipant @chesgreenuk]

….Stephen Wolfram, who gives an explanation of what is going on under the covers.

I read the article last night and it was quite eye opening….. https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/

A number of times the author made comments such as ….

What determines this structure? Ultimately it’s presumably some “neural net encoding” of features of human language. But as of now, what those features might be is quite unknown. In effect, we’re “opening up the brain of ChatGPT” (or at least GPT-2) and discovering, yes, it’s complicated in there, and we don’t understand it—even though in the end it’s producing recognizable human language.

He was hinting, I think, that ChatGPT possibly operates in a similar way to how the human brain might function in some areas.

Ches

[Author]

Posts